Sierra buys Takeoff: long-horizon agents are a runtime
Sierra acquired Takeoff, a 14-month-old long-horizon AI agent company, and is building Horizon. The scarce part isn't the model — it's the state.
Bret Taylor's Sierra just bought a company that is barely a year old and, by its founder's own account, ran with three people. On July 23 Sierra announced it is acquiring Takeoff, a long-horizon AI agent platform, and folding it into a new product called Horizon. The interesting part isn't the price — terms weren't disclosed. It's what a company with Sierra's engineering bench decided it was faster to buy than build.
What actually happened
Takeoff builds agents that run end-to-end jobs across voice, SMS, email, web portals and APIs — work that plays out over days, not one chat session. Founder Aakash Thumaty writes that the company went from zero at the start of 2026 to near eight figures in annual recurring revenue in roughly 14 months, on multiple seven-figure contracts in lending and healthcare. It raised a $15M round from Matrix and Lachy Groom in June 2026, one month before the acquisition.
Sierra had already shipped Horizon on July 16 — agents that "orchestrate outbound and inbound interactions over days or weeks, not just single conversations," with a context engine and long-horizon planning underneath. Sierra prices on outcomes rather than tokens. The two teams are now building Horizon together, which reads less like a talent grab and more like Sierra buying a runtime it wanted shipping this quarter.
Why it matters for your business
Strip away the branding and there's a useful signal here about where agent difficulty actually lives. Sierra can call any frontier model it wants. What it paid for was the unglamorous layer: the thing that holds state across a two-week collections sequence, resumes after a customer ghosts you for six days, remembers what was promised on Tuesday's call when the email lands Friday, and knows the difference between "no answer" and "done."
That's a distributed systems problem wearing an AI costume. It's queues, idempotency, retries, escalation rules, and an audit trail — the same machinery you'd need if the actor were a human temp. If your own agent pilot fell over, it probably wasn't the model reasoning badly. It was that nothing owned the state between steps.
Two things follow for an operator. First, when you evaluate an agent vendor, stop demoing the conversation and start asking about the runtime: where does state live, can you export it, what happens on a partial failure at step seven of eleven, and who gets paged. Second, note the shape of this market — a three-person company reached eight figures in 14 months by bolting agents onto one painful revenue workflow, then got absorbed. Your vendor's independence has a short half-life. Own the workflow definition and the data; rent the runtime, and know what it costs to swap it.
Key takeaways
- Sierra acquired Takeoff on July 23; terms undisclosed, teams now building a joint platform called Horizon
- Takeoff reported going from zero to near eight-figure ARR in about 14 months with a three-person team, on lending and healthcare contracts
- Horizon agents run interactions over days or weeks with a context engine, not single conversations
- The scarce capability being bought is the runtime — state, resumption, escalation — not the model
- Evaluate agent vendors on state ownership, partial-failure behavior, and exportability, not demo quality
Your agent pilot died between steps, not inside them. We build the boring layer — queues, retries, escalation, audit trail — around whatever model you're using, and you own it. See how we build it or tell us where it broke.
Sources: Sierra, Sierra: Horizon, CMSWire.
- #ai-agents
- #acquisitions
- #automation
- #workflow
- #customer-operations
Tommy Rush — Founder, Rush Commerce
Operator turned builder. 15+ years running operations — now shipping the systems businesses run on. More
Get The Rush Report weekly — one email, zero fluff.
Keep reading
Pilot Protocol's $4.5M: a directory isn't a standard
A seed-stage AI agent network wants to be where agents discover each other. Useful, but know the difference between an open spec and someone else's front door.
Read itNvidia backs Safe Superintelligence, discloses no terms
Nvidia's Safe Superintelligence partnership names no dollar figure and no term length. Reported at $5B. Here's how to read AI deals that omit the numbers.
Read it